Merge branch 'main' into fix/nonzero

CITATION.cff

@@ -0,0 +1,34 @@
+cff-version: 1.2.0
+message: "If you use this software, please cite it as below."
+authors:
+  - family-names: "Simard"
+    given-names: "Nathaniel"
+    email: "[email protected]"
+  - family-names: "Fortier-Dubois"
+    given-names: "Louis"
+    email: "[email protected]"
+  - family-names: "Tadjibaev"
+    given-names: "Dilshod"
+    email: "[email protected]"
+  - family-names: "Lagrange"
+    given-names: "Guillaume"
+    email: "[email protected]"
+  - name: "Burn Framework Contributors"
+title: "Burn"
+version: 0.14.0
+date-released: 2024-08-27
+url: "https://burn.dev/"
+repository-code: "https://github.com/tracel-ai/burn"
+license:
+  - MIT
+  - Apache-2.0
+abstract: "Burn is a new comprehensive dynamic Deep Learning Framework built using Rust with extreme flexibility, compute efficiency and portability as its primary goals."
+keywords:
+  - scientific-computing
+  - deep-learning
+  - machine-learning
+  - neural-networks
+  - rust
+  - high-performance-computing
+  - portability
+  - compute-efficiency

Cargo.toml

@@ -151,8 +151,8 @@ systemstat = "0.2.3"
 portable-atomic-util = { version = "0.2.2", features = ["alloc"] }
 
 ### For the main burn branch. ###
-cubecl = { git = "https://github.com/tracel-ai/cubecl", default-features = false, rev = "446dc854ef3f7f93fe4ab2df0f1ab779a7d6a98e" }
-cubecl-common = { git = "https://github.com/tracel-ai/cubecl", default-features = false, rev = "446dc854ef3f7f93fe4ab2df0f1ab779a7d6a98e" }
+cubecl = { git = "https://github.com/tracel-ai/cubecl", default-features = false, rev = "b0fa39bad9d5ecbd98a955e6f7f94c855b6e9a73" }
+cubecl-common = { git = "https://github.com/tracel-ai/cubecl", default-features = false, rev = "b0fa39bad9d5ecbd98a955e6f7f94c855b6e9a73" }
 ### For local development. ###
 # cubecl = { path = "../cubecl/crates/cubecl" }
 # cubecl-common = { path = "../cubecl/crates/cubecl-common" }

backend-comparison/Cargo.toml

@@ -24,14 +24,13 @@ tch-cpu = ["burn/tch"]
 tch-gpu = ["burn/tch"]
 wgpu = ["burn/wgpu", "burn/autotune"]
 wgpu-fusion = ["wgpu", "burn/fusion"]
-# cuda-jit = ["burn-cuda"]
+cuda-jit = ["burn/cuda-jit"]
 
 [dependencies]
 arboard = { workspace = true }
 burn = { path = "../crates/burn", default-features = false }
 burn-common = { path = "../crates/burn-common", version = "0.15.0" }
 burn-wgpu = { path = "../crates/burn-wgpu", default-features = false, version = "0.15.0", optional = true }
-# burn-cuda = { path = "../crates/burn-cuda", version = "0.15.0", optional = true }
 clap = { workspace = true }
 colored = { workspace = true }
 derive-new = { workspace = true }

backend-comparison/src/lib.rs

@@ -129,7 +129,7 @@ macro_rules! bench_on_backend {
 
         #[cfg(feature = "cuda-jit")]
         {
-            use burn_cuda::{Cuda, CudaDevice};
+            use burn::backend::cuda_jit::{Cuda, CudaDevice};
 
             bench::<Cuda>(&CudaDevice::default(), feature_name, url, token);
         }

contributor-book/src/getting-started/configuring-your-editor.md

@@ -23,17 +23,8 @@ To use the debugger, follow these steps:
 1. Open `Command Palette` with `Ctrl+Shift+P` or `F1` and type
    `LLDB: Generate Launch Configurations from Cargo.toml` then select it, this will generate a file
    that should be saved as `.vscode/launch.json`.
-2. Select the configuration from the "run and debug" side panel, then select the target from the
-   list.
-   You may also want to enable debugging by creating a `.vscode/settings.json` file:
-    ```
-    {
-      "rust-analyzer.runnables.extraEnv": {
-        "CARGO_PROFILE_DEV_DEBUG": true
-      }
-    }
-    ```
-    since this repo has `debug = 0` in the root `Cargo.toml` to speed up compilation.
+2. Select the configuration from the "run and debug" side panel, then select the target from the list.
+   Since this repo has `debug = 0` in the root `Cargo.toml` to speed up compilation, you need replace it with `debug = true` in the root `Cargo.toml` when using a debugger and breakpoints with `launch.json` settings.
 3. Now you can enable breakpoints on code through IDE then start debugging the library/binary you
    want, like in the following example:
 

crates/burn-core/src/nn/loss/huber.rs

@@ -37,7 +37,7 @@ impl HuberLossConfig {
 /// The loss for each element of the residuals `r = targets - predictions` is given by
 ///
 /// ```text
-/// L(r) = 0.5 * x^2                  if |r| <= d
+/// L(r) = 0.5 * r^2                  if |r| <= d
 /// L(r) = 0.5 * d^2 + d * (|r| - d)  if |r| >  d
 /// ```
 ///

crates/burn-dataset/Cargo.toml

@@ -21,7 +21,7 @@ fake = ["dep:fake"]
 sqlite = ["__sqlite-shared", "dep:rusqlite"]
 sqlite-bundled = ["__sqlite-shared", "rusqlite/bundled"]
 
-vision = ["dep:flate2", "dep:globwalk", "dep:burn-common"]
+vision = ["dep:flate2", "dep:globwalk", "dep:burn-common", "dep:image"]
 
 # internal
 __sqlite-shared = [

crates/burn-import/onnx-tests/tests/test_onnx.rs

@@ -964,7 +964,6 @@ mod tests {
     }
 
     #[test]
-    #[ignore = "https://github.com/tracel-ai/burn/issues/2080"]
     fn resize_with_scales_1d_linear() {
         // Initialize the model without weights (because the exported file does not contain them)
         let device = Default::default();
@@ -980,10 +979,11 @@ mod tests {
         // The scales are 1.5
         let output = model.forward(input);
 
-        let output_sum = output.sum().into_scalar();
-        let expected_sum = -4.568_224; // from pytorch
-
-        assert!(expected_sum.approx_eq(output_sum, (1.0e-4, 2)));
+        Tensor::<Backend, 3>::from([[[
+            1.5410, 0.3945, -0.7648, -1.9431, -0.8052, 0.3618, -0.6713, -1.2023, -1.3986,
+        ]]])
+        .to_data()
+        .assert_approx_eq(&output.into_data(), 3);
     }
 
     #[test]

crates/burn-jit/src/kernel/interpolate/bicubic.rs

@@ -100,6 +100,7 @@ impl<E: JitElement> InterpolateBicubicShader<E> {
 
         cpa!(scope, input_height = input_shape_2 - 1u32);
         cpa!(scope, output_height = output_shape_2 - 1u32);
+        cpa!(scope, output_height = max(output_height, 1u32));
         cpa!(scope, numerator = h * input_height);
         cpa!(scope, numerator_float = cast(numerator));
         cpa!(scope, output_height_float = cast(output_height));
@@ -129,6 +130,7 @@ impl<E: JitElement> InterpolateBicubicShader<E> {
 
         cpa!(scope, input_width = input_shape_3 - 1u32);
         cpa!(scope, output_width = output_shape_3 - 1u32);
+        cpa!(scope, output_width = max(output_width, 1u32));
         cpa!(scope, numerator = w * input_width);
         cpa!(scope, numerator_float = cast(numerator));
         cpa!(scope, output_width_float = cast(output_width));

crates/burn-jit/src/kernel/interpolate/bilinear.rs

@@ -101,6 +101,7 @@ impl InterpolateBilinearShader {
 
         cpa!(scope, numerator_int = input_shape_2 - 1u32);
         cpa!(scope, denominator_int = output_shape_2 - 1u32);
+        cpa!(scope, denominator_int = max(denominator_int, 1u32));
         cpa!(scope, factor_float = cast(h));
         cpa!(scope, numerator_float = cast(numerator_int));
         cpa!(scope, denominator_float = cast(denominator_int));
@@ -115,6 +116,7 @@ impl InterpolateBilinearShader {
 
         cpa!(scope, numerator_int = input_shape_3 - 1u32);
         cpa!(scope, denominator_int = output_shape_3 - 1u32);
+        cpa!(scope, denominator_int = max(denominator_int, 1u32));
         cpa!(scope, factor_float = cast(w));
         cpa!(scope, numerator_float = cast(numerator_int));
         cpa!(scope, denominator_float = cast(denominator_int));

crates/burn-ndarray/src/ops/interpolate.rs

@@ -96,8 +96,8 @@ pub(crate) fn bilinear_interpolate<E: FloatNdArrayElement>(
     let (batch_size, channels, in_height, in_width) = x.dim();
     let [out_height, out_width] = output_size;
 
-    let y_ratio = ((in_height - 1) as f64) / ((out_height - 1) as f64);
-    let x_ratio = ((in_width - 1) as f64) / ((out_width - 1) as f64);
+    let y_ratio = ((in_height - 1) as f64) / (core::cmp::max(out_height - 1, 1) as f64);
+    let x_ratio = ((in_width - 1) as f64) / (core::cmp::max(out_width - 1, 1) as f64);
 
     let out_element_num = batch_size * channels * out_height * out_width;
     let strides = (
@@ -174,8 +174,8 @@ pub(crate) fn bicubic_interpolate<E: FloatNdArrayElement>(
     let (batch_size, channels, in_height, in_width) = x.dim();
     let [out_height, out_width] = output_size;
 
-    let y_ratio = ((in_height - 1) as f64) / ((out_height - 1) as f64);
-    let x_ratio = ((in_width - 1) as f64) / ((out_width - 1) as f64);
+    let y_ratio = ((in_height - 1) as f64) / (core::cmp::max(out_height - 1, 1) as f64);
+    let x_ratio = ((in_width - 1) as f64) / (core::cmp::max(out_width - 1, 1) as f64);
 
     let out_element_num = batch_size * channels * out_height * out_width;
     let strides = (

crates/burn-tensor/src/tests/module/bicubic_interpolate.rs

@@ -86,7 +86,6 @@ mod tests {
     }
 
     #[test]
-    #[ignore = "https://github.com/tracel-ai/burn/issues/2080"]
     fn test_1d_bicubic() {
         // Initialize the model without weights (because the exported file does not contain them)
         let device = Default::default();

crates/burn-tensor/src/tests/module/bilinear_interpolate.rs

@@ -86,7 +86,6 @@ mod tests {
     }
 
     #[test]
-    #[ignore = "https://github.com/tracel-ai/burn/issues/2080"]
     fn test_1d_bilinear() {
         // Initialize the model without weights (because the exported file does not contain them)
         let device = Default::default();

crates/burn-train/src/renderer/tui/progress.rs

@@ -86,7 +86,7 @@ impl ProgressBarView {
 
         let iteration = Gauge::default()
             .gauge_style(Style::default().fg(Color::Yellow))
-            .ratio(self.progress);
+            .ratio(self.progress.min(1.0));
         let eta = Paragraph::new(Line::from(vec![
             Span::from(" ("),
             Span::from(self.eta).italic(),